Summary
On September 10, 2026, Abacus.AI launched Smaug, three open-weight models fine-tuned for agentic use: Smaug Agentic (a ~2T Kimi K3 fine-tune for complex coding loops), Smaug Flash (a DeepSeek-based model for personal agents across WhatsApp, Telegram, and Slack), and Smaug Mini (small and further-tunable for chatbots). Abacus says its Smaug fine-tuning technique improves long-running agent loops 15-20% without added cost; the models are on Hugging Face and via its RouteLLM API.
What changed
Abacus.AI published three open-weight Smaug models on September 10, 2026 on Hugging Face (and via RouteLLM), positioning them as cost-efficient, self-hostable replacements for frontier models on agentic workloads.
Why it matters
Open-weight models specialized for long-running agent loops let enterprises run agents on their own GPUs at a claimed 10-100x lower cost than frontier APIs, with data kept in-house — a direct pressure point on Anthropic and OpenAI pricing for high-volume agent traffic.
Evidence excerpt
Smaug is a fine-tuning technique that improves the performance of long-running agentic loops by 15-20% without increasing cost; three models are available for download on Hugging Face and through the RouteLLM API.